Write a prompt, pick a document type and a primary organisation. Tick extra orgs if this is a cross-org document. SDI drafts it using the house principles, the leading documents of the scoped orgs, and up to five reference documents you pick by id.

• The house principles (platform + active org) prepended to the system message.
• The top-3 leading documents for each scope org (truncated to ~2 KB each).
• Any reference documents you picked explicitly (truncated to ~3 KB each).
• The optional public-policy footnote, if you ticked the box.

Generate shows a preview with token counts and the ids of leading docs it cited. If it looks right, hit Save as draft: SDI inserts a documents row, commits the markdown to SDI-data via PR, and tags the new doc as awaiting_approval. Discard simply throws the preview away — nothing is saved.

The Agent is a conversation with read access to your corpus via tools. Each session is scoped to one organisation (narrow) or open to all your orgs (wide). Tools the agent uses:

• search_documents — full-text search, supports filters (type, lang, only_authoritative).
• semantic_search — dense-vector search (needs embeddings; see Insights & Mining).
• get_document — fetches the full markdown of a doc by id.
• list_events — assemblies, deadlines, external obligations in a date range.
• aggregate_entities — top people / organisations / laws / places in the corpus.

Every session keeps its user prompts, assistant replies, and — for audit — the tool calls and their results. Platform admins can read any session. Delete a session to remove its messages entirely.

• Use Generate when you know the shape you want (a position paper, an opinion, an analysis).
• Use Agent when you're exploring: "what did we decide about X?", "who is mentioned across these minutes?".